Image compression; Face recognition

Travel Credentials, require facial photographs to becompressed to approximately one kilobyte, posing significantchallenges for automated face recognition. This papersystematically evaluates six lossy compression algorithms(JPEG, JPEG2000, JPEG-AI, JPEG-FzT, JPEG-XL,and WebP) on ten face recognition models (seven opensourceand three proprietary configurations) at two resolutions(112×112 and 224×224 pixels). The results indicatethat JPEG-AI and WebP consistently yield the lowestdegradation of recognition across models and operatingpoints; on the size-compliant subset at 224×224 pixels,JPEG-AI is moreover statistically superior to all remainingcodecs. JPEG-AI and JPEG-FzT are the only algorithmscompressing all 224×224 images to the 1 kB target,making them uniquely suitable when a strict per-imagesize guarantee is required, whereas JPEG2000 performspoorest. The reported codec ranking remains stable acrossdemographic subgroups, acquisition conditions, cross-agepairs, compressed-to-compressed matching, and two additionalstate-of-the-art recognition models.

Travel Credentials, require facial photographs to becompressed to approximately one kilobyte, posing significantchallenges for automated face recognition. This papersystematically evaluates six lossy compression algorithms(JPEG, JPEG2000, JPEG-AI, JPEG-FzT, JPEG-XL,and WebP) on ten face recognition models (seven opensourceand three proprietary configurations) at two resolutions(112×112 and 224×224 pixels). The results indicatethat JPEG-AI and WebP consistently yield the lowestdegradation of recognition across models and operatingpoints; on the size-compliant subset at 224×224 pixels,JPEG-AI is moreover statistically superior to all remainingcodecs. JPEG-AI and JPEG-FzT are the only algorithmscompressing all 224×224 images to the 1 kB target,making them uniquely suitable when a strict per-imagesize guarantee is required, whereas JPEG2000 performspoorest. The reported codec ranking remains stable acrossdemographic subgroups, acquisition conditions, cross-agepairs, compressed-to-compressed matching, and two additionalstate-of-the-art recognition models.